🧩

CEVA

💡
Thesis. Licensing hit a three-year high with real operating leverage behind it — margin from 3% to 11% — and a major AI platform company has licensed NeuPro-M for custom silicon. But royalties grew 1% on 16% device growth, so the flywheel the whole thesis depends on has not started turning.
Reviewed 11 September 2026, after Q2 results. Next review after Q3 earnings on 10 November 2026.
⚠️
Scope note. This is edge AI — phones, earbuds, cars, sensors — not data-centre AI. CEVA's demand driver is consumer device volume, not hyperscaler capex. It sits in this file as a deliberate single rather than a constituent of the AI infrastructure build-out, and should be read that way.

What it does

CEVA licenses silicon IP — chip design blueprints. It doesn't manufacture anything.
Like Arm, the money comes twice: an upfront licence fee, then a royalty on every chip shipped containing the design.
What it licenses: DSPs (digital signal processors, which handle audio, vision and radio), NeuPro NPUs (neural processing units that run AI on the device itself rather than in the cloud), and wireless connectivity blocks for Bluetooth, Wi-Fi, UWB and cellular IoT. CEO is Amir Panush. Around 406 people, 327 of them engineers.
The pitch is "Physical AI" — connectivity, sensing and inference converging inside everyday devices. Whether that pitch pays depends entirely on royalties, which is where the problem is.

Bull case

  • Licensing at a three-year high. Q2 2026, reported 10 August: licensing revenue $18.2M, up 21%. Total revenue $29.0M, up 13%. Trailing-twelve-month licensing $69.6M
  • Ten agreements signed in the quarter, including two first-time customers and two signed directly with OEMs rather than chipmakers — CEVA being chosen at product level, not by a chip vendor
  • A major platform company picked NeuPro-M. A leading global AI and computing platform company licensed it as the foundation of a custom AI silicon programme
  • The operating leverage is real. Non-GAAP operating margin went from 3% to 11%, with operating expenses up only 3% while revenue grew 13%. GAAP operating loss narrowed from $4.5M to $2.1M
  • Guidance raised to 13–15% growth for FY26 from 12%, with non-GAAP operating income guided up roughly 70%
  • 87% GAAP gross margin — the economics of selling blueprints
  • Volume is there. 567M devices shipped, up 16%. Wi-Fi up 28%, cellular IoT a record 68M
  • The mix is shifting. AI was over 20% of licensing revenue in 2025, with more than a dozen NeuPro wins to date. Bluetooth High Data Throughput carries a higher royalty rate and ramps from late 2026

Bear case

  • Royalties grew 1%. On 16% device growth. This is the recurring half of the business and it is flat — the single most important fact on this page
  • The royalty take is tiny. 567M devices produced $10.8M, roughly two cents per device. A company with real pricing power over its licensees would see volume growth flow through
  • Still loses money on a GAAP basis. Q2 GAAP net loss $2.9M, EPS −$0.10. The non-GAAP $0.08 adds back about $5.2M, mostly stock compensation, which dilutes holders whatever the accounting says
  • The market noticed. The stock fell roughly 10% on a beat-and-raise, which tells you what investors are actually watching
  • Parts of the base are shrinking — Bluetooth and industrial IoT unit shipments declined
  • Licensing is lumpy by nature. A three-year high is a quarter, not a run rate
  • A $58.5M equity raise (3,000,000 shares at $19.50) diluted holders
  • It competes with Arm, Synopsys and Cadence at a fraction of their scale

Major customers

CEVA does not disclose named customers or concentration percentages. What it does disclose are classes of licensee:
  • Semiconductor companies building CEVA IP into their own chips — the traditional base
  • OEMs licensing directly — two signed in Q2. A meaningful shift, because it means the device maker chose CEVA rather than inheriting it from a chip vendor
  • A leading global AI and computing platform company — the NeuPro-M custom silicon programme, unnamed
  • A leading PC OEM — a strategic NPU licensing agreement signed in 2025
  • Smartphone, consumer IoT, automotive and industrial device makers — the royalty base, 567M devices in the quarter
🔗
Read-through — limited, and that is the point. CEVA's demand comes from smartphone and consumer IoT unit volumes, not hyperscaler capex. It has no meaningful supply-chain connection to anything else in this file: no data centre, no GPUs, no optics.
Its closest structural comparable is
🧩
ARM — the same fee-plus-royalty IP licensing model, at vastly greater scale and with far deeper ecosystem lock-in. The comparison is useful mainly for what it shows about royalty economics: Arm's royalty rate per chip is rising, and CEVA's is two cents and flat.

What would change the view

  • Royalty revenue growth — currently +1%. This single number decides the thesis
  • Whether licensing holds near the three-year high or reverts to the lumpy average
  • Non-GAAP operating margin progressing from 11%
  • NPU royalties beginning to contribute — expected once 2026 designs reach silicon in 2027
  • Bluetooth High Data Throughput ramp through 2027–28, where the higher royalty rate lives
  • GAAP profitability, still negative

Update log

  • 11 Sep 2026 — Reviewed. No change to the thesis or the verdict.
  • 14 Aug 2026 — Full review after Q2 2026 results. Licensing strength and the NeuPro-M platform win added to the bull case; flat royalty growth flagged as the dominant risk.
Research and education only — not investment advice.